KeKsBoTer
commented
3 weeks ago Thanks for your interest! I tuned the hyperparameters for the datasets in the paper. Your scene might be different (e.g.larger) and require different values.
First, try do adjust which gaussian get quantized with the color_importance_include and gaussian_importance_include parameters. If you decrease these values, the compression will be less aggressive, but give better results.
You can also try to increase the codebook sizes with --gaussian_codebook_size <size> and --color_codebook_size <size>.
Great work overall! This is a really great start to having 3DGS models be more portable for the web. However, I've encountered some issues while testing this compression technique with custom data - my own .ply files-
The .ply splat model rendering:
The .npz compression model rendering:
Not sure I understand what went wrong - cloned the repo and ran the command against the .ply file.
Did you guys use more iterations than the hard-coded defaults to produce the results shown in the examples?After testing, it seems like running more finetune iterations yields worse results in terms of the visual representation.